- inability to transparently explain decisions of llm
- making it hard to understand how llm process inputs and generate outputs
- this creates challenges in debugging, trust, and accountability
- developers often rely on trial-and-error or additional tools to interpret their behavior